Table 5 Mean (standard deviation) of AUC for the Gleason 3 vs. Gleason 4 classification problem across the four hold-one-site-out folds with \(F{S}_{sd}^{\theta ,\kappa }\) and \(F{S}_{d}^{\theta ,\kappa }\) \(\theta \in \{SFS,WLCX,mRMR,ROC\}\), \(\kappa \in \{LDA,QDA,SVM,RF\}\).

For each classifier model, the top 5 most stable and discriminating or most discriminating features were employed for constructing \(F{S}_{sd}\) and \(F{S}_{d}\) respectively. For every feature selection-classification pair four models were trained and validated, one model for every possible combination of three of the four sites. The three chosen sites were combined and used for training and the held out site was used for validation. The improvement between \(F{S}_{sd}\) and \(F{S}_{d}\) is shown. A positive improvement indicates that \(F{S}_{sd}\) outperformed \(F{S}_{d}\). Improvement in \(F{S}_{sd}\) over \(F{S}_{d}\) occurred in 13 of the 16 cases, with the average improvement in those 13 scenarios being 5.92% compared to 2.28% when \(F{S}_{d}\) was superior compared to \(F{S}_{sd}\).

Quick links

Search